Research On The Related Issues Of Melody Extraction

Melody extraction is an important issue in music information retrieval(MIR).The purpose of melody extraction is to convert the waveform file into a sequence of human voice fundamental frequencies.As an important acoustic feature,fundamental frequency plays an important role in many fields.Therefore,the main melody extraction plays an important role in the interaction of music information.The tradition main melody extraction method mainly includes two modules,one is the candidate fundamental frequency extraction and the other is the main fundamental frequency discrimination.At present,the candidate fundamental frequency extraction module mainly has the problem that it extracts so many false fundamental frequencies.The main fundamental frequency discrimination module mainly has two disadvantages,one is insufficient utilization of short-time features in time frame and the other is the reliance on extraction accuracy of the initial phase of the melody.This paper made improvements to the above drawbacks,and the main innovation includes the following two parts:1.This paper proposed a candidate fundamental frequency extraction module base on tandem filter system.The system consists of three parts,preprocessing part,Hyper-Fourier transform filter part and the transverse stripe filter part.First,there exists so much interfere in the original signal without preprocessing,so we use the repeatability of the music,and apply the state-of-art source separation method to process the original signal.Then we use Hyper-Fourier Transform to filter the accompaniment spectrum components.This part reduce the number of non-human voice fundamental frequency in all candidate fundamental frequencies.Finally,we believe that the STFT spectrogram have the original information of the melody frequency structure,here we used a transverse stripe filter system to eliminate the non-possible fundamental frequency position.2.This paper proposed a machine learning classifier base on Mel Frequency Cepstral Coefficient,Harmonic Feature,Spectral Shape Feature,Spectrum Contrast Feature to decision which frame is human voice frame.The experiment was conducted on two terms,single feature and combination feature.Finally,we found a feature combination that both perform well on frame precision and frame recall rate.